Cluster-Based Search Thresholds for Accurate Person Matching
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing person search technologies in crowded spaces face issues with erroneous detection and search omission due to varying optimal thresholds for different outer appearances and image-taking conditions, leading to inadequate search performance.
Innovation Solution
A search device that derives thresholds for each cluster of feature values obtained through clustering, using a threshold deriving unit to identify appropriate thresholds for search features within a feature database, enabling accurate person identification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a single threshold is used for all camera combinations, then the search process is simple and fast, but search accuracy deteriorates due to varying optimal thresholds for different outer appearances and image-taking conditions
Solution Approach 1:
The patent segments the feature value space into multiple clusters based on outer appearance characteristics (e.g., clothing color, pattern) and image-taking conditions. Each cluster is assigned a dedicated threshold, replacing the single threshold approach. This segmentation enables the system to select appropriate thresholds for different person types and conditions, significantly improving search accuracy while maintaining manageable complexity through automated cluster-based threshold selection.
2Measurement precision
If thresholds are adjusted for each camera combination, then measurement precision improves, but the system complexity and processing time increase
Solution Approach 1:
The patent changes the threshold parameter dynamically based on the cluster to which the extracted feature values belong. Instead of using fixed thresholds for each camera combination, the system adjusts the threshold parameter according to the detected outer appearance and imaging conditions. This parameter change approach allows precise threshold selection for different scenarios without managing complex camera-specific threshold tables, reducing system complexity while maintaining high measurement precision.
3Reliability
If a high threshold is used to reduce erroneous detection, then false positives decrease, but search omission increases for persons with distinctive appearances
Solution Approach 1:
The patent applies local quality by assigning different threshold values to different clusters of feature values. Each cluster, representing a specific outer appearance type or imaging condition, receives a customized threshold optimized for its characteristics. This allows the system to use higher thresholds for clusters with stable feature distributions (reducing false positives) while using lower thresholds for clusters with high variability (preventing search omission). The local optimization of threshold quality for each cluster simultaneously improves both detection reliability and search completeness.
Data Source
AI summary
By taking each of a plurality of clusters obtained by clustering a plurality of feature values stored in a feature database (48) as a target cluster, a threshold deriving unit (47) derives a threshold for the target cluster from a distribution of the feature values in the target cluster. By using, as a target threshold, a threshold for a cluster to which a search feature, which is a feature value for an image in a search request, belongs among the plurality of clusters, a search unit (44) identifies a feature value corresponding to the search feature from the plurality of feature values stored in the feature database (48).